Compare commits

...

7 Commits

Author SHA1 Message Date
rohitvinodmalhotra@gmail.com
532a284d5c migrate to use OH version 2025-01-26 15:24:35 -05:00
rohitvinodmalhotra@gmail.com
43f6104967 Merge branch 'main' into eval/visualcodebench 2025-01-26 15:14:28 -05:00
openhands
e249b920ff feat: adapt Design2Code block detection for in-memory evaluation 2024-11-30 19:28:22 +00:00
rohitvinodmalhotra@gmail.com
d920a69f69 adding back server code 2024-11-30 14:00:25 -05:00
openhands
a8ce888981 refactor: adapt Design2Code evaluation metrics 2024-11-30 17:17:05 +00:00
rohitvinodmalhotra@gmail.com
e22ddc0dd6 uncomment agent run 2024-11-26 17:00:07 -05:00
rohitvinodmalhotra@gmail.com
c370912f12 adding eval scripts 2024-11-26 16:57:19 -05:00
7 changed files with 1246 additions and 0 deletions

1
.gitignore vendored
View File

@@ -176,6 +176,7 @@ evaluation/gorilla/data
evaluation/toolqa/data evaluation/toolqa/data
evaluation/scienceagentbench/benchmark evaluation/scienceagentbench/benchmark
evaluation/commit0_bench/repos evaluation/commit0_bench/repos
evaluation/visualcodebench/
# openhands resolver # openhands resolver
output/ output/

View File

@@ -0,0 +1,674 @@
from collections import Counter
from copy import deepcopy
from difflib import SequenceMatcher
from io import BytesIO
from bs4 import BeautifulSoup, Comment, NavigableString, Tag
import cv2
import numpy as np
import torch
from colormath.color_conversions import convert_color
from colormath.color_diff import delta_e_cie2000
from colormath.color_objects import LabColor, sRGBColor
from PIL import Image, ImageChops, ImageColor
from scipy.optimize import linear_sum_assignment
from transformers import CLIPModel, CLIPProcessor
from openhands.core.logger import openhands_logger as logger
def calculate_similarity(block1, block2):
"""Calculate text similarity between two blocks using SequenceMatcher."""
text_similarity = SequenceMatcher(None, block1['text'], block2['text']).ratio()
return text_similarity
def adjust_cost_for_context(cost_matrix, consecutive_bonus=1.0, window_size=20):
"""Adjust cost matrix by considering context similarity."""
if window_size <= 0:
return cost_matrix
n, m = cost_matrix.shape
adjusted_cost_matrix = np.copy(cost_matrix)
for i in range(n):
for j in range(m):
if adjusted_cost_matrix[i][j] >= -0.5:
continue
nearby_matrix = cost_matrix[
max(0, i - window_size) : min(n, i + window_size + 1),
max(0, j - window_size) : min(m, j + window_size + 1),
]
flattened_array = nearby_matrix.flatten()
sorted_array = np.sort(flattened_array)[::-1]
sorted_array = np.delete(
sorted_array, np.where(sorted_array == cost_matrix[i, j])[0][0]
)
top_k_elements = sorted_array[-window_size * 2 :]
bonus = consecutive_bonus * np.sum(top_k_elements)
adjusted_cost_matrix[i][j] += bonus
return adjusted_cost_matrix
def create_cost_matrix(A, B):
"""Create cost matrix for block matching."""
n = len(A)
m = len(B)
cost_matrix = np.zeros((n, m))
for i in range(n):
for j in range(m):
cost_matrix[i, j] = -calculate_similarity(A[i], B[j])
return cost_matrix
def calculate_distance_max_1d(x1, y1, x2, y2):
"""Calculate maximum 1D distance between points."""
return max(abs(x2 - x1), abs(y2 - y1))
def calculate_ratio(h1, h2):
"""Calculate ratio between two heights."""
return max(h1, h2) / min(h1, h2)
def rgb_to_lab(rgb):
"""Convert RGB color to Lab color space."""
rgb_color = sRGBColor(rgb[0], rgb[1], rgb[2], is_upscaled=True)
lab_color = convert_color(rgb_color, LabColor)
return lab_color
def color_similarity_ciede2000(rgb1, rgb2):
"""Calculate color similarity using CIEDE2000 formula."""
lab1 = rgb_to_lab(rgb1)
lab2 = rgb_to_lab(rgb2)
delta_e = delta_e_cie2000(lab1, lab2)
similarity = max(0, 1 - (delta_e / 100))
return similarity
def merge_blocks_wo_check(block1, block2):
"""Merge two blocks without additional checks."""
merged_text = block1['text'] + ' ' + block2['text']
x_min = min(block1['bbox'][0], block2['bbox'][0])
y_min = min(block1['bbox'][1], block2['bbox'][1])
x_max = max(
block1['bbox'][0] + block1['bbox'][2], block2['bbox'][0] + block2['bbox'][2]
)
y_max = max(
block1['bbox'][1] + block1['bbox'][3], block2['bbox'][1] + block2['bbox'][3]
)
merged_bbox = (x_min, y_min, x_max - x_min, y_max - y_min)
merged_color = tuple(
(color1 + color2) // 2
for color1, color2 in zip(block1['color'], block2['color'])
)
return {'text': merged_text, 'bbox': merged_bbox, 'color': merged_color}
def find_maximum_matching(A, B, consecutive_bonus, window_size):
"""Find maximum matching between two sets of blocks."""
cost_matrix = create_cost_matrix(A, B)
cost_matrix = adjust_cost_for_context(cost_matrix, consecutive_bonus, window_size)
row_ind, col_ind = linear_sum_assignment(cost_matrix)
current_cost = cost_matrix[row_ind, col_ind].tolist()
return list(zip(row_ind, col_ind)), current_cost, cost_matrix
def remove_indices(lst, indices):
"""Remove indices from list in reverse order."""
for index in sorted(indices, reverse=True):
if index < len(lst):
lst.pop(index)
return lst
def merge_blocks_by_list(blocks, merge_list):
"""Merge blocks according to merge list."""
pop_list = []
while merge_list:
i = merge_list[0][0]
j = merge_list[0][1]
blocks[i] = merge_blocks_wo_check(blocks[i], blocks[j])
pop_list.append(j)
merge_list.pop(0)
if merge_list:
new_merge_list = []
for k in range(len(merge_list)):
if (
merge_list[k][0] != i
and merge_list[k][1] != i
and merge_list[k][0] != j
and merge_list[k][1] != j
):
new_merge_list.append(merge_list[k])
merge_list = new_merge_list
remove_indices(blocks, pop_list)
return blocks
def difference_of_means(list1, list2):
"""Calculate difference of means between two lists."""
counter1 = Counter(list1)
counter2 = Counter(list2)
for element in set(list1) & set(list2):
common_count = min(counter1[element], counter2[element])
counter1[element] -= common_count
counter2[element] -= common_count
unique_list1 = [item for item in counter1.elements()]
unique_list2 = [item for item in counter2.elements()]
mean_list1 = sum(unique_list1) / len(unique_list1) if unique_list1 else 0
mean_list2 = sum(unique_list2) / len(unique_list2) if unique_list2 else 0
if mean_list1 - mean_list2 > 0:
if min(unique_list1) > min(unique_list2):
return mean_list1 - mean_list2
return 0.0
return mean_list1 - mean_list2
def find_possible_merge(A, B, consecutive_bonus, window_size, debug=False):
"""Find possible merges between blocks."""
merge_bonus = 0.0
merge_windows = 1
def sortFn(value):
return value[2]
while True:
A_changed = False
B_changed = False
matching, current_cost, cost_matrix = find_maximum_matching(
A, B, merge_bonus, merge_windows
)
if len(A) >= 2:
merge_list = []
for i in range(len(A) - 1):
new_A = deepcopy(A)
new_A[i] = merge_blocks_wo_check(new_A[i], new_A[i + 1])
new_A.pop(i + 1)
updated_matching, updated_cost, _ = find_maximum_matching(
new_A, B, merge_bonus, merge_windows
)
diff = difference_of_means(current_cost, updated_cost)
if diff > 0.05:
merge_list.append([i, i + 1, diff])
merge_list.sort(key=sortFn, reverse=True)
if merge_list:
A_changed = True
A = merge_blocks_by_list(A, merge_list)
matching, current_cost, cost_matrix = find_maximum_matching(
A, B, merge_bonus, merge_windows
)
if len(B) >= 2:
merge_list = []
for i in range(len(B) - 1):
new_B = deepcopy(B)
new_B[i] = merge_blocks_wo_check(new_B[i], new_B[i + 1])
new_B.pop(i + 1)
updated_matching, updated_cost, _ = find_maximum_matching(
A, new_B, merge_bonus, merge_windows
)
diff = difference_of_means(current_cost, updated_cost)
if diff > 0.05:
merge_list.append([i, i + 1, diff])
merge_list.sort(key=sortFn, reverse=True)
if merge_list:
B_changed = True
B = merge_blocks_by_list(B, merge_list)
matching, current_cost, cost_matrix = find_maximum_matching(
A, B, merge_bonus, merge_windows
)
if not A_changed and not B_changed:
break
matching, _, _ = find_maximum_matching(A, B, consecutive_bonus, window_size)
return A, B, matching
def merge_blocks_by_bbox(blocks):
"""Merge blocks with same bounding box."""
merged_blocks = {}
for block in blocks:
bbox = tuple(block['bbox'])
if bbox in merged_blocks:
existing_block = merged_blocks[bbox]
existing_block['text'] += ' ' + block['text']
existing_block['color'] = [
(ec + c) / 2 for ec, c in zip(existing_block['color'], block['color'])
]
else:
merged_blocks[bbox] = block
return list(merged_blocks.values())
def mask_bounding_boxes_with_inpainting(image, bounding_boxes):
"""Mask bounding boxes in image using inpainting."""
image_cv = cv2.cvtColor(np.array(image), cv2.COLOR_RGB2BGR)
mask = np.zeros(image_cv.shape[:2], dtype=np.uint8)
height, width = image_cv.shape[:2]
for bbox in bounding_boxes:
x_ratio, y_ratio, w_ratio, h_ratio = bbox
x = int(x_ratio * width)
y = int(y_ratio * height)
w = int(w_ratio * width)
h = int(h_ratio * height)
mask[y : y + h, x : x + w] = 255
inpainted_image = cv2.inpaint(image_cv, mask, 3, cv2.INPAINT_TELEA)
return Image.fromarray(cv2.cvtColor(inpainted_image, cv2.COLOR_BGR2RGB))
def rescale_and_mask(image, blocks):
"""Rescale image and mask blocks."""
if blocks:
image = mask_bounding_boxes_with_inpainting(image, blocks)
width, height = image.size
if width < height:
new_size = (width, width)
else:
new_size = (height, height)
return image.resize(new_size, Image.LANCZOS)
def calculate_clip_similarity(image1, image2, blocks1, blocks2):
"""Calculate CLIP similarity between two images."""
model = CLIPModel.from_pretrained('openai/clip-vit-base-patch32')
processor = CLIPProcessor.from_pretrained('openai/clip-vit-base-patch32')
device = 'cuda' if torch.cuda.is_available() else 'cpu'
model = model.to(device)
# Mask and preprocess images
image1_masked = rescale_and_mask(image1, [block['bbox'] for block in blocks1])
image2_masked = rescale_and_mask(image2, [block['bbox'] for block in blocks2])
inputs = processor(
images=[image1_masked, image2_masked], return_tensors='pt', padding=True
)
inputs = {k: v.to(device) for k, v in inputs.items()}
# Calculate features and similarity
with torch.no_grad():
image_features = model.get_image_features(**inputs)
image_features1 = image_features[0].unsqueeze(0)
image_features2 = image_features[1].unsqueeze(0)
image_features1 /= image_features1.norm(dim=-1, keepdim=True)
image_features2 /= image_features2.norm(dim=-1, keepdim=True)
similarity = (image_features1 @ image_features2.T).item()
return similarity
def rgb_to_hex(rgb):
"""Convert an RGB tuple to hexadecimal format."""
return '{:02X}{:02X}{:02X}'.format(*rgb)
class ColorPool:
def __init__(self, offset=0):
color_values = list(range(10, 251, 16))
color_list = [((r + offset) % 256, (g + offset) % 256, (b + offset) % 256)
for r in color_values for g in color_values for b in color_values]
self.color_pool = [rgb_to_hex(color) for color in color_list]
def pop_color(self):
if self.color_pool:
return self.color_pool.pop()
else:
raise NotImplementedError
def process_html_str(html_str, offset=0):
"""Process HTML string to assign unique colors to text elements."""
soup = BeautifulSoup(html_str, 'html.parser')
def update_style(element, property_name, value):
important_value = f"{value} !important"
styles = element.attrs.get('style', '').split(';')
updated_styles = [s for s in styles if not s.strip().startswith(property_name) and len(s.strip()) > 0]
updated_styles.append(f"{property_name}: {important_value}")
element['style'] = '; '.join(updated_styles).strip()
# Set background color of all elements to transparent white
for element in soup.find_all(True):
update_style(element, 'background-color', 'rgba(255, 255, 255, 0.0)')
color_pool = ColorPool(offset)
text_tags = ['p', 'h1', 'h2', 'h3', 'h4', 'h5', 'h6', 'div', 'span', 'a', 'b', 'li',
'table', 'td', 'th', 'button', 'footer', 'header', 'figcaption']
for tag in soup.find_all(text_tags):
color = f"#{color_pool.pop_color()}"
update_style(tag, 'color', color)
update_style(tag, 'opacity', '1.0')
return str(soup)
def similar(n1, n2):
"""Check if two numbers are similar within a threshold."""
return abs(n1 - n2) <= 8
def find_different_pixels(image1, image2):
"""Find pixels that differ between two images."""
if image1.size != image2.size:
logger.warning("Images are not the same size")
return None
image1 = image1.convert('RGB')
image2 = image2.convert('RGB')
pixels1 = image1.load()
pixels2 = image2.load()
different_pixels = []
for x in range(image1.size[0]):
for y in range(image1.size[1]):
r1, g1, b1 = pixels1[x, y]
r2, g2, b2 = pixels2[x, y]
if similar((r1 + 50) % 256, r2) and similar((g1 + 50) % 256, g2) and similar((b1 + 50) % 256, b2):
different_pixels.append((y, x))
return np.stack(different_pixels) if different_pixels else None
def extract_text_with_color(html_str):
"""Extract text and color information from HTML string."""
def get_color(tag):
if 'style' in tag.attrs:
styles = tag['style'].split(';')
color_style = [s for s in styles if 'color' in s and 'background-color' not in s]
if color_style:
color = color_style[-1].split(':')[1].strip().replace(" !important", "")
if color[0] == "#":
return color
else:
try:
if color.startswith('rgb'):
color = tuple(map(int, color[4:-1].split(',')))
else:
color = ImageColor.getrgb(color)
return '#{:02x}{:02x}{:02x}'.format(*color)
except ValueError:
logger.warning(f"Unable to identify or convert color: {color}")
return None
return None
def extract_text_recursive(element, parent_color='#000000'):
if isinstance(element, Comment):
return None
elif isinstance(element, NavigableString):
text = element.strip()
return (text, parent_color) if text else None
elif isinstance(element, Tag):
current_color = get_color(element) or parent_color
children_texts = filter(None, [extract_text_recursive(child, current_color)
for child in element.children])
return list(children_texts)
soup = BeautifulSoup(html_str, 'html.parser')
body = soup.body
return extract_text_recursive(body) if body else []
def flatten_tree(tree):
"""Flatten a nested tree structure into a list."""
flat_list = []
def flatten(node):
if isinstance(node, list):
for item in node:
flatten(item)
else:
flat_list.append(node)
flatten(tree)
return flat_list
def get_blocks_from_image_diff_pixels(image, html_text_color_tree, different_pixels):
"""Extract text blocks from image using color differences."""
image_cv = cv2.cvtColor(np.array(image), cv2.COLOR_RGB2BGR)
x_w = image_cv.shape[0]
y_w = image_cv.shape[1]
def hex_to_bgr(hex_color):
hex_color = hex_color.lstrip('#')
rgb = tuple(int(hex_color[i:i+2], 16) for i in (0, 2, 4))
return rgb[::-1]
def get_intersect(arr1, arr2):
arr1_reshaped = arr1.view([('', arr1.dtype)] * arr1.shape[1])
arr2_reshaped = arr2.view([('', arr2.dtype)] * arr2.shape[1])
common_rows = np.intersect1d(arr1_reshaped, arr2_reshaped)
return common_rows.view(arr1.dtype).reshape(-1, arr1.shape[1])
blocks = []
for item in html_text_color_tree:
try:
color = np.array(hex_to_bgr(item[1]), dtype="uint8")
except:
continue
lower = color - 4
upper = color + 4
mask = cv2.inRange(image_cv, lower, upper)
coords = np.column_stack(np.where(mask > 0))
coords = get_intersect(coords, different_pixels)
if coords.size == 0:
continue
x_min, y_min = np.min(coords, axis=0)
x_max, y_max = np.max(coords, axis=0)
# Get average color from original image
color_coords = coords.copy()
color_coords = color_coords[color_coords[:, 0] <= x_max]
color_coords = color_coords[color_coords[:, 1] <= y_max]
colors = [image_cv[x, y] for x, y in color_coords]
avg_color = tuple(map(int, np.mean(colors, axis=0)))[::-1] # Convert BGR to RGB
blocks.append({
'text': item[0].lower(),
'bbox': (y_min / y_w, x_min / x_w, (y_max - y_min + 1) / y_w, (x_max - x_min + 1) / x_w),
'color': avg_color
})
return blocks
def get_blocks_from_html(html_str, image1):
"""Extract text blocks from HTML and image."""
# Process HTML with two different color offsets
html_str_1 = process_html_str(html_str, offset=0)
html_str_2 = process_html_str(html_str, offset=50)
# Render both HTML versions to images
# TODO: Screenshot html_str_2
filter_color = (255, 0, 0)
image2 = Image.new("RGB", image1.size, filter_color)
# Find pixels that differ between the two rendered images
different_pixels = find_different_pixels(image1, image2)
if different_pixels is None:
logger.warning("Unable to get pixels with different colors")
return []
# Extract text and color information from HTML
html_text_color_tree = flatten_tree(extract_text_with_color(html_str_1))
try:
blocks = get_blocks_from_image_diff_pixels(image1, html_text_color_tree, different_pixels)
except Exception as e:
logger.warning(f"Unable to get blocks: {e}")
return []
return blocks
def evaluate(task, generated_img):
"""Evaluate generated image against reference image using multiple metrics."""
# Load reference image
post_image = task['post_image']
# Extract blocks from HTML and images
post_blocks = get_blocks_from_html(task['post_html'], post_image)
gen_blocks = get_blocks_from_html(task['gen_html'], generated_img)
print("block details", post_blocks, gen_blocks)
if not post_blocks or not gen_blocks:
# Fallback to basic CLIP and pixel comparison if no blocks available
clip_score = calculate_clip_similarity(post_image, generated_img, [], [])
logger.info(f'CLIP similarity score: {clip_score}')
# Pixel comparison
diff = ImageChops.difference(generated_img, post_image)
pixel_match = not diff.getbbox()
logger.info(
f"Pixel difference analysis: {'No difference' if pixel_match else 'Differences found'}"
)
return clip_score > 0.95 or pixel_match
# Merge blocks with same bounding boxes
post_blocks = merge_blocks_by_bbox(post_blocks)
gen_blocks = merge_blocks_by_bbox(gen_blocks)
# Find optimal block matching
consecutive_bonus, window_size = 0.1, 1
gen_blocks_m, post_blocks_m, matching = find_possible_merge(
gen_blocks, deepcopy(post_blocks), consecutive_bonus, window_size
)
# Filter matches with low similarity
filtered_matching = []
for i, j in matching:
text_similarity = calculate_similarity(gen_blocks_m[i], post_blocks_m[j])
if text_similarity >= 0.5:
filtered_matching.append([i, j, text_similarity])
matching = filtered_matching
if not matching:
logger.warning('No matching blocks found')
clip_score = calculate_clip_similarity(
post_image, generated_img, gen_blocks, post_blocks
)
return clip_score > 0.95
# Calculate metrics for matched blocks
indices1 = [item[0] for item in matching]
indices2 = [item[1] for item in matching]
# Calculate unmatched areas
unmatched_area_1 = sum(
block['bbox'][2] * block['bbox'][3]
for i, block in enumerate(gen_blocks_m)
if i not in indices1
)
unmatched_area_2 = sum(
block['bbox'][2] * block['bbox'][3]
for j, block in enumerate(post_blocks_m)
if j not in indices2
)
total_unmatched_area = unmatched_area_1 + unmatched_area_2
# Calculate metrics for matched blocks
matched_areas = []
text_scores = []
position_scores = []
color_scores = []
for i, j, text_similarity in matching:
# Area
block_area = (
gen_blocks_m[i]['bbox'][2] * gen_blocks_m[i]['bbox'][3]
+ post_blocks_m[j]['bbox'][2] * post_blocks_m[j]['bbox'][3]
)
matched_areas.append(block_area)
# Position similarity
position_similarity = 1 - calculate_distance_max_1d(
gen_blocks_m[i]['bbox'][0] + gen_blocks_m[i]['bbox'][2] / 2,
gen_blocks_m[i]['bbox'][1] + gen_blocks_m[i]['bbox'][3] / 2,
post_blocks_m[j]['bbox'][0] + post_blocks_m[j]['bbox'][2] / 2,
post_blocks_m[j]['bbox'][1] + post_blocks_m[j]['bbox'][3] / 2,
)
# Color similarity
color_similarity = color_similarity_ciede2000(
gen_blocks_m[i]['color'], post_blocks_m[j]['color']
)
text_scores.append(text_similarity)
position_scores.append(position_similarity)
color_scores.append(color_similarity)
# Calculate final scores
total_area = sum(matched_areas) + total_unmatched_area
size_score = sum(matched_areas) / total_area if total_area > 0 else 0
text_score = np.mean(text_scores) if text_scores else 0
position_score = np.mean(position_scores) if position_scores else 0
color_score = np.mean(color_scores) if color_scores else 0
clip_score = calculate_clip_similarity(
post_image, generated_img, gen_blocks, post_blocks
)
# Combine scores with equal weights
final_score = 0.2 * (
size_score + text_score + position_score + color_score + clip_score
)
logger.info('Evaluation scores:')
logger.info(f'- Size score: {size_score:.3f}')
logger.info(f'- Text score: {text_score:.3f}')
logger.info(f'- Position score: {position_score:.3f}')
logger.info(f'- Color score: {color_score:.3f}')
logger.info(f'- CLIP score: {clip_score:.3f}')
logger.info(f'- Final score: {final_score:.3f}')
return final_score > 0.8 # Consider it a match if final score > 80%
def png_to_bytes(png):
buffer = BytesIO()
png.save(buffer, format='PNG')
image_bytes = buffer.getvalue()
return image_bytes
def bytes_to_image(image_bytes):
"""Convert bytes to a Pillow Image object."""
return Image.open(BytesIO(image_bytes))
if __name__ == '__main__':
first_image = Image.open('./evaluation/visualcodebench/data/1/post.png')
image = Image.open('./evaluation/visualcodebench/data/1/prev.png')
html_file = open('./evaluation/visualcodebench/data/1/post/index.html', 'r')
first_html = html_file.read()
html_file.close()
html_file = open('./evaluation/visualcodebench/data/1/prev/index.html', 'r')
gen_html = html_file.read()
html_file.close()
sample = {'post_image': first_image, "post_html": first_html, "gen_html": gen_html}
evaluate(sample, image)

View File

@@ -0,0 +1,97 @@
import base64
import os
from io import BytesIO
import pandas as pd
from huggingface_hub import snapshot_download
from PIL import PngImagePlugin
from tqdm import tqdm
from openhands.core.logger import openhands_logger as logger
REPO_DOWNLOAD_DIR = (
'./evaluation/visualcodebench/' # Directory to store the downloaded repository
)
def download_repository():
"""
Download the entire repository from Hugging Face Hub.
This function clones the repository into REPO_DOWNLOAD_DIR.
"""
repo_id = 'rvmalhot/VisualCodeBench'
try:
logger.info(f"Downloading repository '{repo_id}'...")
snapshot_download(
repo_id=repo_id,
local_dir=REPO_DOWNLOAD_DIR,
repo_type='dataset',
ignore_patterns=None, # Download all files
)
logger.info(f"Repository downloaded to '{REPO_DOWNLOAD_DIR}'.")
except Exception as e:
logger.error(f"Error downloading repository '{repo_id}': {e}")
raise e
def format_task_dict(example):
instance_id = example['id']
prev_remote_path = os.path.join(REPO_DOWNLOAD_DIR, f'data/{instance_id}/prev')
post_remote_path = os.path.join(REPO_DOWNLOAD_DIR, f'data/{instance_id}/post')
# Check if 'prev' and 'post' directories exist
prev_exists = os.path.exists(prev_remote_path)
post_exists = os.path.exists(post_remote_path)
if prev_exists and post_exists:
skip = False
else:
skip = True
task = {
'instance_id': instance_id,
'prev_image': example['prev_image'],
'post_image': example['post_image'],
'changes': example['changes'],
'prev_code_files': example['prev_code_files'],
'post_code_files': example['post_code_files'],
'skip': skip,
}
return task
def prepare_visualcodebench(dataset):
logger.info('Processing dataset')
dataset_processed = []
for example in tqdm(dataset['train']):
formatted_example = format_task_dict(example)
if formatted_example['skip']:
continue
del formatted_example['skip']
dataset_processed.append(formatted_example)
return pd.DataFrame(dataset_processed)
def pil_image_to_base64(image: PngImagePlugin.PngImageFile) -> str:
"""
Converts a PIL image to a Base64-encoded string.
Parameters:
- image (PngImagePlugin.PngImageFile): The PIL image to convert.
Returns:
- str: The Base64-encoded string of the image.
"""
if not isinstance(image, PngImagePlugin.PngImageFile):
raise ValueError(
'The provided image is not a PIL.PngImagePlugin.PngImageFile instance.'
)
buffered = BytesIO()
image.save(buffered, format='PNG')
img_bytes = buffered.getvalue()
img_base64 = base64.b64encode(img_bytes).decode('utf-8')
base64_with_prefix = f'data:image/png;base64,{img_base64}'
return [base64_with_prefix]

View File

@@ -0,0 +1,247 @@
# FILE: run_infer.py
import asyncio
import os
import shutil
import tempfile
from functools import partial
import pandas as pd
from datasets import load_dataset
# from evaluation.benchmarks.visualcodebench.eval import capture_screenshot
from evaluation.benchmarks.visualcodebench.prepare import (
REPO_DOWNLOAD_DIR,
download_repository,
pil_image_to_base64,
prepare_visualcodebench,
)
from evaluation.utils.shared import (
EvalMetadata,
assert_and_raise,
codeact_user_response,
make_metadata,
prepare_dataset,
reset_logger_for_multiprocessing,
run_evaluation,
)
from openhands.controller.state.state import State
from openhands.core.config import (
AppConfig,
SandboxConfig,
get_llm_config_arg,
)
from openhands.core.config.utils import parse_arguments
from openhands.core.logger import openhands_logger as logger # Import OpenHands logger
from openhands.core.main import create_runtime, run_controller
from openhands.events.action.commands import CmdRunAction
from openhands.events.action.message import MessageAction
from openhands.events.observation.commands import CmdOutputObservation
from openhands.runtime.base import Runtime
from openhands.utils.async_utils import call_async_from_sync
# Define workspace and output directories
WORKSPACE_DIR = './workspace'
FAKE_RESPONSES = {
'CodeActAgent': partial(codeact_user_response, encapsulate_solution=True),
}
def get_config(
metadata: EvalMetadata,
) -> AppConfig:
config = AppConfig(
default_agent=metadata.agent_class,
run_as_openhands=False,
runtime='eventstream',
max_iterations=metadata.max_iterations,
sandbox=SandboxConfig(
base_container_image='python:3.12-bookworm',
enable_auto_lint=True,
use_host_network=False,
),
# do not mount workspace
workspace_base=None,
workspace_mount_path=None,
)
config.set_llm_config(metadata.llm_config)
return config
def initialize_runtime(
runtime: Runtime,
instance: pd.Series, # this argument is not required
):
"""Initialize the runtime for the agent.
This function is called before the runtime is used to run the agent.
"""
logger.info('-' * 30)
logger.info('BEGIN Runtime Initialization Fn')
logger.info('-' * 30)
workspace_dir_name = instance['instance_id']
obs: CmdOutputObservation
action = CmdRunAction(command='mkdir -p /workspace/{workspace_dir_name}')
action.timeout = 600
logger.info(action, extra={'msg_type': 'ACTION'})
obs = runtime.run_action(action)
logger.info(obs, extra={'msg_type': 'OBSERVATION'})
assert_and_raise(
obs.exit_code == 0,
f'Failed to create /workspace/{workspace_dir_name}: {str(obs)}',
)
file_path = REPO_DOWNLOAD_DIR + f'data/{workspace_dir_name}/prev/index.html'
runtime.copy_to(file_path, f'/workspace/{workspace_dir_name}')
logger.info(f'Copied code file for instance {workspace_dir_name}')
action = CmdRunAction(command=f'cd /workspace/{workspace_dir_name}')
action.timeout = 600
logger.info(action, extra={'msg_type': 'ACTION'})
obs = runtime.run_action(action)
logger.info(obs, extra={'msg_type': 'OBSERVATION'})
assert_and_raise(
obs.exit_code == 0,
f'Failed to cd to /workspace/{workspace_dir_name}: {str(obs)}',
)
logger.info('-' * 30)
logger.info('END Runtime Initialization Fn')
logger.info('-' * 30)
def complete_runtime(
runtime: Runtime,
instance: pd.Series, # this argument is not required, but it is used to get the workspace_dir_name
) -> str:
# TODO: extract edited HTML file from agent workspace
# temp_zip = runtime.copy_from(f'/workspace/{instance.instance_id}')
# file_name = f'/workspace/{instance.instance_id}/index.html'
# with zipfile.ZipFile(temp_zip, 'r') as zip_ref:
# if file_name in zip_ref.namelist():
# with zip_ref.open(file_name) as file:
# file_content = file.read().decode('utf-8') # Decode bytes to string
# else:
# raise FileNotFoundError(f"'{file_name}' not found in the ZIP archive.")
with tempfile.TemporaryDirectory() as tmpdir:
src_folder = REPO_DOWNLOAD_DIR + f'data/{instance.instance_id}/post/'
shutil.copytree(src_folder, tmpdir, dirs_exist_ok=True)
# image = capture_screenshot(tmpdir)
# if image is not None:
# shutil.copy(os.path.join(tmpdir, 'final_screenshot.png'), REPO_DOWNLOAD_DIR)
def process_instance(
instance: pd.Series, metadata: EvalMetadata, reset_logger: bool = True
):
config = get_config(metadata)
# Setup the logger properly, so you can run multi-processing to parallelize the evaluation
if reset_logger:
log_dir = os.path.join(metadata.eval_output_dir, 'infer_logs')
reset_logger_for_multiprocessing(logger, instance.instance_id, log_dir)
else:
logger.info(f'Starting evaluation for instance {instance.instance_id}.')
# =============================================
# build instruction
# =============================================
# Prepare instruction
instruction = (
f"Modify the HTML/CSS according to the following instruction:\n\n"
f"{instance['changes']}\n\n"
)
instruction += (
'IMPORTANT: You should ONLY interact with the environment provided '
'to you AND NEVER ASK FOR HUMAN HELP.\n'
)
# =============================================
# create sandbox and run the agent
# =============================================
runtime: Runtime = create_runtime(config)
call_async_from_sync(runtime.connect)
try:
initialize_runtime(runtime, instance=instance)
image_urls = pil_image_to_base64(instance['prev_image'])
action = MessageAction(content=instruction, image_urls=image_urls)
state: State | None = asyncio.run(
run_controller(
config=config,
initial_user_action=action,
runtime=runtime,
fake_user_response_fn=FAKE_RESPONSES[metadata.agent_class],
)
)
if state is None:
raise ValueError('State should not be None.')
# =============================================
# result evaluation
# =============================================
return_val = complete_runtime(runtime, instance)
logger.info(f'Return value {return_val}')
finally:
runtime.close()
# TODO: return EVAL output
def main():
"""Main function to run the evaluation."""
# args = parse_args()
args = parse_arguments()
logger.info(f"\n{'='*80}\nStarting VisualCodeBench Evaluation\n{'='*80}")
logger.info(f'Agent: {args.agent_cls}')
logger.info(f'Model: {args.llm_config}')
logger.info(f'Max iterations: {args.max_iterations}')
logger.info(f'Eval limit: {args.eval_n_limit}')
logger.info(f'Num workers: {args.eval_num_workers}\n')
logger.info(f'Eval output: {args.eval_output_dir}\n')
# Step 1: Download the entire repository once
logger.info('Downloading repository...')
download_repository()
# Step 2: Load Dataset
logger.info('Loading dataset...')
dataset = load_dataset(REPO_DOWNLOAD_DIR)
# Step 3: Prepare dataset
llm_config = get_llm_config_arg(args.llm_config)
if llm_config is None:
logger.error(f'Could not find LLM config: {args.llm_config}')
raise ValueError(f'Could not find LLM config: {args.llm_config}')
metadata = make_metadata(
llm_config,
'VisualCodeBench',
args.agent_cls,
args.max_iterations,
args.eval_note,
'evaluation/output/',
)
output_file = os.path.join(metadata.eval_output_dir, 'output.jsonl')
dataset = prepare_visualcodebench(dataset)
instances = prepare_dataset(dataset, output_file, eval_n_limit=args.eval_n_limit)
# Step 4: Run eval
run_evaluation(
instances, metadata, output_file, args.eval_num_workers, process_instance
)
if __name__ == '__main__':
main()

View File

@@ -0,0 +1,46 @@
#!/bin/bash
set -eo pipefail
source "evaluation/utils/version_control.sh"
# Check if required arguments are provided
if [ "$#" -lt 4 ]; then
echo "Usage: $0 [model_config] [commit_hash] [agent_cls] [eval_limit] [num_workers]"
echo "Example: $0 llm.eval_gpt_4o_mini HEAD CodeActAgent 5 1"
exit 1
fi
MODEL_CONFIG=$1
COMMIT_HASH=$2
AGENT_CLS=$3
EVAL_LIMIT=$4
NUM_WORKERS=${5:-1} # Default to 1 worker if not specified
# Checkout the specified commit
checkout_eval_branch
if [ -z "$AGENT" ]; then
echo "Agent not specified, use default CodeActAgent"
AGENT="CodeActAgent"
fi
get_openhands_version
echo "AGENT: $AGENT"
echo "OPENHANDS_VERSION: $OPENHANDS_VERSION"
echo "MODEL_CONFIG: $MODEL_CONFIG"
COMMAND="export PYTHONPATH=evaluation/benchmarks/visualcodebench:\$PYTHONPATH && poetry run python evaluation/benchmarks/visualcodebench/run_infer.py \
--agent-cls $AGENT \
--llm-config $MODEL_CONFIG \
--max-iterations 5 \
--eval-num-workers $NUM_WORKERS \
--eval-note $OPENHANDS_VERSION" \
if [ -n "$EVAL_LIMIT" ]; then
echo "EVAL_LIMIT: $EVAL_LIMIT"
COMMAND="$COMMAND --eval-n-limit $EVAL_LIMIT"
fi
# Run the command
eval $COMMAND

View File

@@ -0,0 +1,167 @@
import http
import os
import socket
import socketserver
import threading
import time
from io import BytesIO
import requests
from PIL import Image, ImageChops
from playwright.sync_api import sync_playwright
from openhands.core.logger import openhands_logger as logger
def get_free_port():
"""Find a free port to run the HTTP server."""
with socket.socket(socket.AF_INET, socket.SOCK_STREAM) as s:
s.bind(('', 0))
return s.getsockname()[1]
def start_http_server(tmpdir):
port = get_free_port()
class CustomHTTPRequestHandler(http.server.SimpleHTTPRequestHandler):
def translate_path(self, path):
# Serve files from the specified directory instead of the current working directory
path = super().translate_path(path)
relative_path = os.path.relpath(path, os.getcwd())
return os.path.join(tmpdir, relative_path)
handler = CustomHTTPRequestHandler
server = socketserver.TCPServer(('', port), handler)
return server, port
def capture_screenshot(tmpdir):
server, port = start_http_server(tmpdir)
server_thread = threading.Thread(target=server.serve_forever)
server_thread.daemon = True
server_thread.start()
time.sleep(10)
image = None
try:
server_url = f'http://localhost:{port}/'
if not is_server_reachable(server_url):
raise RuntimeError(f'Server not reachable at {server_url}')
screenshot_path = os.path.join(tmpdir, 'final_screenshot.png')
capture_screenshot_playwright(server_url, screenshot_path)
image = Image.open(screenshot_path)
image.load()
finally:
# Shut down the server and clean up
server.shutdown()
server.server_close()
return image
def is_server_reachable(url):
"""
Check if the local server is reachable.
"""
try:
response = requests.get(url, timeout=5) # Set a 5-second timeout
if response.status_code == 200:
logger.info(f'Server is reachable at {url}')
return True
else:
logger.warning(
f'Server responded with status code {response.status_code} at {url}'
)
return False
except requests.ConnectionError as e:
logger.error(f'Failed to connect to server at {url}: {e}')
return False
def capture_screenshot_playwright(url, screenshot_path):
"""Capture a screenshot of the given URL using Playwright."""
try:
with sync_playwright() as p:
logger.info('Launching browser...')
browser = p.chromium.launch(timeout=10000) # 10 seconds for browser launch
logger.info('Creating a new page...')
page = browser.new_page()
logger.info(f'Navigating to URL: {url}')
try:
page.goto(url, timeout=60 * 1000) # Set timeout to 5 seconds
logger.info('Page navigation completed.')
except Exception as e:
logger.warning(f'Page navigation timed out. {e}. Continuing...')
logger.info('Waiting for network to be idle...')
try:
page.wait_for_load_state(
'networkidle', timeout=60 * 1000
) # Set timeout to 5 seconds
logger.info('Page load state reached.')
except Exception as e:
logger.warning(f'Page load state timed out. {e}. Continuing...')
logger.info('Capturing screenshot...')
page.screenshot(
path=screenshot_path, full_page=True
) # Capture full page screenshot
logger.info(f'Screenshot saved to {screenshot_path}')
browser.close()
return True
except Exception as e:
logger.error(f'Error capturing screenshot with Playwright: {e}')
return False
def evaluate(task, screenshot_path):
"""Compare generated screenshot with post_image using CLIP score."""
try:
import torch
from transformers import CLIPModel, CLIPProcessor
# Load CLIP model and processor
model = CLIPModel.from_pretrained('openai/clip-vit-base-patch32')
processor = CLIPProcessor.from_pretrained('openai/clip-vit-base-patch32')
# Load images
post_image = Image.open(BytesIO(task['post_image']))
generated_img = Image.open(screenshot_path)
# Process images
inputs = processor(
images=[post_image, generated_img], return_tensors='pt', padding=True
)
# Get image features
image_features = model.get_image_features(**inputs)
# Calculate cosine similarity
similarity = torch.nn.functional.cosine_similarity(
image_features[0].unsqueeze(0), image_features[1].unsqueeze(0)
).item()
logger.info(f'CLIP similarity score: {similarity}')
return similarity > 0.95 # Consider it a match if similarity > 95%
except Exception as e:
logger.error(f'Error in CLIP evaluation: {e}')
# Fallback to pixel comparison if CLIP fails
try:
post_image = Image.open(BytesIO(task['post_image']))
generated_img = Image.open(screenshot_path)
# Compare images directly without converting to bytes
diff = ImageChops.difference(generated_img, post_image)
logger.info(
f"Pixel difference analysis: {'No difference' if not diff.getbbox() else 'Differences found'}"
)
return not diff.getbbox()
except Exception as ex:
logger.error(f'Error in fallback evaluation: {ex}')
return False

14
poetry.lock generated
View File

@@ -1108,6 +1108,20 @@ humanfriendly = ">=9.1"
[package.extras] [package.extras]
cron = ["capturer (>=2.4)"] cron = ["capturer (>=2.4)"]
[[package]]
name = "colormath"
version = "3.0.0"
description = "Color math and conversion library."
optional = false
python-versions = "*"
files = [
{file = "colormath-3.0.0.tar.gz", hash = "sha256:3d4605af344527da0e4f9f504fad7ddbebda35322c566a6c72e28edb1ff31217"},
]
[package.dependencies]
networkx = ">=2.0"
numpy = "*"
[[package]] [[package]]
name = "comm" name = "comm"
version = "0.2.2" version = "0.2.2"